Admission Control Of Network Service Systems Based On Stochastic Optimization

After entering the21st century, the technology of computer networks has gained great progress, and the Internet has brought great convenience for peoples life and work. As the demand for network services is continually increasing, the number and the scale of network service systems have also been growing constantly. The network applica-tions and services, such as instant communication and video on demand etc, have been widely used by humans. In terms of network structure, the single service node network system represented by traditional C/S structure has been gradually replaced by distribut-ed network service system. The P2P (Peer-to-Peer) and CDN (Content Delivery Net-work) overlay network structures are two most popular network, and have been applied by most network service systems currently. Although the network service systems es-tablished on distributed structure have obtained quite good performance and scalability, they still face some performance bottleneck in case of the increasing user scale. And since the management strategies of many distributed network system are inefficient, the resource utilization of these system is usually low, which means the system operator is difficult to obtain the ideal benefit. As the basis of whether a request can be admitted, admission control policy can obviously influence the load and benefit of a network ser-vice system. Thus, the optimization of admission control policy in a distributed network service system has significant value in theory and practical applications.This dissertation focuses on the problem of admission control in a distributed net-work system, including the admission control processes of user quests in the service node of system. Using the methodology of stochastic optimization, this dissertation gives the dynamic evolution models by mathematical method to describe the operation process of system with the consideration of practical network background. The op-timization algorithm is applied to obtain the optimal admission control policy on the basis of the presented model. The main works and contributions of this dissertation are listed in the following part.Firstly, for the admission control problem of a distributed network service system, this dissertation gives an optimization method on the basis of the semi-Markov deci-sion process (SMDP). To solve the optimal admission control policy for the system with fixed parameters, a SMDP model is proposed to describe the admission control process at each service node. When a new request arrives at a service node, the controller de-cides whether to admit it according to the system state and request type. Then, to deal with the updating of admission control policy when the system parameters are changing over time, an admission control policy switching mechanism is designed. With the poli- cy switching mechanism, the system chooses an admission control policy from the basic admission control policy set when the change of system parameters reaches a certain degree. A new SMDP model is proposed to describe the policy switching control pro-cess of the system, and the action is chosen according to the request arrival rates at each service node and the time epoch. The optimization object is to maximize the long run benefit of system. The parameterized random policy and the deterministic policy are designed as the policy of two SMDPs respectively. The gradient-based method and the Q-learning method are applied to solve the optimal admission control policy with fixed parameters and the optimal switching control rule for policy switching control process. The proposed stochastic optimization models can give a quite accurate description of system, and can also reflect the influences of the factors in the system.Secondly, consider the admission control characteristics of practical network sys-tems, an event-based optimization method is proposed. Using the ideal of event, this dissertation gives a method to describe the control process of a practical system with events. Then, considering the policy optimization problem of a system when the event-based parameterized random policy is chosen, the sensitivity-based analysis method is used to derive the system performance difference equation and performance gradient equation with respect to the policy parameter. With the method of stochastic approxi-mation, the online performance gradient estimation algorithm and policy iteration algo-rithm are proposed. Further more, two types of constraints are considered in the policy optimization. Based on the event, and together with the Lagrange approach, a poli-cy optimization algorithm is proposed to solve the optimal policy with constraint. In comparison with the MDP-based optimization, the event-based optimization can better describe the system characteristics, and reduce the algorithm complexity effectively in the optimization process.Lastly, an event-based optimization method is proposed to deal with the admis-sion control problem with probabilistic constraint of admitted request been dropped. With the event-based method, the admission control process of the system is described by different events include a request is admitted and served, a request is admitted but dropped, a request is rejected. Then, this dissertation gives the definitions of risk event and risk value of an event, and derived the risk value of system. The risk value of sys-tem in the event-based model is the long run probability of an admitted request been dropped, which converts the original probabilistic constraint to general constraint. By applying the Lagrange approach, the Lagrange reward and performance of system is given, and with the event-based optimization algorithms, the optimal admission control policy which satisfies the constraints is obtained.